2013 European Modelling Symposium
Median Filter Approach for Removal of Baseline Wander
in Photoplethysmography Signals
Adewale Emmanuel Awodeyi Stephen R. Alty
Dept. of Engineering and Design Dept. of Engineering and Design
London South Bank University London South Bank University
London, United Kingdom London, United Kingdom
e-mail: awodeyia@lsbu.ac.uk e-mail: steve.alty@lsbu.ac.uk
Abstract— Removal of baseline wander is a crucial step in the
signal conditioning stage of photoplethysmography signals.
Hence, a method for removing the baseline wander from
photoplethysmography based on two-stages of median filtering
is proposed in this paper. Recordings from Physionet database
are used to validate the proposed method. In this paper, the
two-stage moving average filtering is also applied to remove
baseline wander in photoplethysmography signals for
comparison with our novel two-stage median filtering method.
Our experiment results show that the performance of two-
stage median filtering method is more effective in removing
baseline wander from photoplethysmography signals. This
median filtering method effectively improves the cross
correlation with minimal distortion of the signal of interest.
Although the method is proposed for baseline wander in
photoplethysmography signals, it can be applied to other
biomedical signals as well.
Keywords- photoplethysmography; cardiovascular; baseline
wander.
I. INTRODUCTION
Photoplethysmography (PPG) is a non-invasive optical
technique for detecting cardiovascular characteristics from
the fingertip while the heart is pumping [1]. It uses an
infrared light source to illuminate the finger on one side, and
a photo detector placed on the other side measures the small
variations in the transmitted or reflected light intensity. The
variations in the photo detector signal are related to changes
in blood volume inside the tissue with heart rate. The plot for
this variation against time is typically referred to as a
photoplethysmographic or PPG signal, but are commonly
contaminated by baseline wander (BW) caused by
respiration at frequencies varying between 0.15 and 0.5 Hz
[2]. The transmittance and reflectance are two basic types of
photoplethysmography. The transmittance or reflectance
PPG is applicable to a restricted body part, such as the
finger or the ear lobe because of the limited penetration
depth of the light through organ tissue. Therefore, it is very
important that the signal is correctly filtered to remove BW
but preserve as far as possible important information. The
median filter is currently enjoying greater theoretical
attention and practical application. Nonetheless, there are yet
relatively few papers published on median filter behavior.
Median filters operate by selecting the middle value of an
ascending-ordered sequence of numbers. These numbers are
taken from a moving window on the data to be processed.
978-1-4799-2578-0/13 $31.00 © 2013 IEEE 248
261
DOI 10.1109/EMS.2013.45
Mohammad Ghavami
Dept. of Engineering and Design
London South Bank University
London, United Kingdom
e-mail: ghavamim@lsbu.ac.uk
The operation can be used in data processing to reject
glitches, align events and to pass step functions as well as
enhance discontinuities in the data. After this brief overview,
the properties of median filtering will be presented in section
IV. The methods and experimental results are described in
section V. The correlation co-efficient is used to measure the
performance of the proposed method.
II. BACKGROUND
In order to develop the necessary sensors to collect
relevant data for further processing and effective
classification algorithms to accurately separate them into
healthy and unhealthy characteristics in biomedical
application of signal processing for assessing cardiovascular
health, we decided to understand anatomy and physiology of
human body and develop a hand-held device to quickly,
accurately and non-invasively assess a patient’s risk of
developing cardiovascular disease in the next ten years. The
human body is made up of many component systems which
include the nervous system, the cardiovascular system and
the musculoskeletal system, among others. Biomedical
application of signal processing for assessing cardiovascular
health have been applied to investigate a host of biological
and physiological systems that manifest themselves as
signal.
III. PPG FEATURES AND PARAMETERS
The quality of the PPG signal depends on artifacts, the
spacing between the pulse waves and the variations in the
stroke volume of the heart can be seen as changes in the
amplitudes of the pulse wave. However, the information has
shown to be useful in diagnosis and in gauging the severity
of illness in newborns. Photoplethysmography signals
consist of a large DC component, which is attributed to the
total blood volume of the examined tissue, and a pulsatile
AC component, which is synchronous to the pumping action
of the heart. The two maxima observed in the PPG waveform
are called systolic and diastolic peaks, and they can provide
valuable information about the cardiovascular system [3].
The time duration between two consecutive systolic peaks
gives the instantaneous heart rate. The AC component is
superimposed onto a large DC component. The DC
component must be removed to measure the AC waveform
with a high signal–to–noise ratio. Since the useful AC signal
is only a very small portion of the whole signal, an effective
amplification circuit is also required to extract desired
information from it. A typical PPG waveform is shown in
the figure below.
median is that the running median filter can be applied
along a given trace, or across adjacent traces. In general,
though, a window of values will be used, a median extracted
and placed on the output trace at the middle of the window,
the window shifted one data point along the sequence, a
median extracted, window shift, etc. While the median filter
may be used to filter a trace in time, it is also often used as a
spatial process [5].

V. METHOD AND EXPERIMENT RESULTS

Figure 1: A typical PPG waveform, as described by [4].
IV. MEDIAN PROPERTIES
As noted previously, the median filter M is a data-
dependent operator which extracts a value from a sequence
of numbers. For example, M (5, 1000, 6, - 1, -1, 7, 9) when
The PPG vector or matrix data extracted from Physionet
database [6] is first filtered before yielding a clean signal
called PPG original signal as shown in Figure 2(a). An
artificial baseline wander (BW) is generated by using a
combination of sine and cosine waves as displayed in Figure
2(b). The frequencies of the sine and cosine waves are 0.4Hz
and 0.2Hz respectively while their corresponding amplitudes
are 60mV and 120mV. Thereby, we have corrupted the PPG
original signal with measured amounts of baseline wander to
generate the PPG test signal as shown in Figure 2(c).
1500
PPG original signal

ordered, becomes M (-1,-1, 5, 6, 7, 9, 1000) = 6. The 1000

inclusion of a large number that deviate from what is normal Amplitude(mV)

in this sequence disturbs the median selection very little but 500

If a sequence is symmetric about some point (for example,

an event has a symmetric noise distribution), then the median 0

and mean selections will be the same [5]. For example, the
-500
sequence M (0, 5, 6, 7, 8, 9, 14) = 7.
The median selection is also invariant to scaling. That is, -1000
0 2 4 6 8 10
if all the numbers in the sequence are scaled or Time(sec)

exponentiated (assuming positive numbers), then a given (a)

number will occupy the same ordered position before and 200
Baseline wander signal
after scaling. For example, if a positive sequence of 150
numbers has a median Xm, and then the sequence values are 100
all squared, the new median will be X2m. The mean does not
Amplitude(mV)

50
have this exponentiation property.
0

-50

-100

-150

It is also a nonlinear operator, as it does not conform to -200

0 2 4 6 8 10
Time(sec)
the definition of linearity:
(b)
2000
For example, take PPG test signal
1500

1000
M = Median filter
Amplitude(mV)

500

-500

-1000

-1500
0 2 4 6 8 10
Time(sec)
Thus the median filter is nonlinear. If M were instead a (c)
mean filter in the above example, then equation (1) would Figure 2: Signal types (a) PPG original signal (b) Baseline wander signal
hold (with 7 = 7). Another interesting property of the (c) PPG test signal.

262
249
 We applied the two-stage median filter (we will refer to performance is measured by root mean square error
this as TSMdF) method to PPG test signal. The TSMdF have (RMSE). The lower RMSE value the better quality of the
variable window sizes for removal of BW in samples with an signal. RMSE can be defined as the square root of the mean
average first stage and second stage of window size set to 0.3 of the square error [9].
and 0.6 respectively. The PPG signal reconstructed after
removing BW is shown in Figure 3(a). As a comparison, a
two-stage moving average filter (TSMaF) is also used to
remove BW in the PPG test signal. Figure 3(b) displays the
result of applying the TSMaF to remove the BW.
1500
PPG signal reconstructed by TSMdF Where yi is the values actually observed, ŷ i is the values
1000 predicted. The RMSE value for PPG signal reconstructed
reach 27.3mV for TSMdF while 41.8mV for TSMaF.
Amplitude(mV)

500

1500
0 PPG original signal
PPG signal reconstructed by TSMdF
-500
1000
-1000
0 2 4 6 8 10
Time(sec)

A m p litu d e (m V )
(a) 500
1500
PPG signal reconstructed by TSMaF

1000
0
Amplitude(mV)

500

0 -500

-500

-1000
-1000
0 2 4 6 8 10 0 2 4 6 8 10
Time(sec)
Time(sec)
(b) Figure 4: The solid red line is the PPG original signal and the dotted line is
Figure 3: Simulation results (a) PPG signal reconstructed by TSMdf (b) the PPG signal reconstructed by the TSMdF method.
PPG signal reconstructed by TSMaF.

However, careful comparison between Figure 3(a) and 1500

3(b) indicate that TSMdf is better at removing BW than PPG original signal
TSMaF. In order to compare the results, we evaluate the PPG signal reconstructed by TSMaF
performance of the correction methods with cross-correlation 1000
[7].
A m p litu d e (m V )

500

Where Xi is each X0 value through X(n-1) of PPG original 0

signal, Yi is each Y0 value through Y(n-1) of the PPG signal
reconstructed. The cross correlation Rxy of the PPG original
and the PPG signal reconstructed reach 0.9989 for TSMdF
correction method while only 0.9975 for TSMaF. The PPG -500
test signal and corresponding PPG signal reconstructed are
shown in Figure (4) and Figure (5) respectively. The
relationship is illustrated graphically using a normalized XY
plot in Matlab [8]. Where two sets of signals are strongly -1000
0 2 4 6 8 10
linked together, we say they have a higher positive Time(sec)
correlation as in Figure 6, while Figure 7 shows a noticeably Figure 5: The solid red line is the PPG original signal and the dotted line is
lower positive correlation. The embedding distortion the PPG signal reconstructed by the TSMaF method.

263
250
 XY Scatter Plot for Original vs. TSMdF XY Scatter Plot for Original vs. TSMaF
1500 1500

y = 0.9961*x + 1.124e-013 y = 0.9836*x + 4.164e-014

1000 1000

500 500

TSMaF
TSMdF

0 0

-500 -500
data 1 data 1
linear linear

-1000 -1000
-1000 -500 0 500 1000 1500 -1000 -500 0 500 1000 1500
PPG Original PPG Original

Figure 6: Scatter plot for TSMdF vs. Original PPG, showing a higher Figure 7: Scatter plot for TSMaF vs. Original PPG, showing a lower
correlation than Figure 7. correlation than Figure 6.

VI. CONCLUSION REFERENCES

The present paper has used various quality of
measurement. In this paper, the results of experiment and
correlation coefficient measure indicate clearly that our
method can give a very satisfactory performance on removal
of baseline wander in photoplethysmography signal. Ten
extracted photoplethysmography signals from Physionet
database were also used to evaluate the performance of the
presented method. Work is underway on cascaded filtering
of photoplethysmography signals to remove other artifacts
commonly found in PPG signals: spikes, power line
interference and noise. Future work will involve peak
detection and heart rate extraction from
photoplethysmography signals and data from various
patients will be analyzed for healthy and unhealthy
classification.
ACKNOWLEDGMENT
The first author would like to express his thanks to
biomedical engineering and communications (BiMEC) for
providing facility to work at London South Bank University
and my supervisor for providing moral support and
encouragement towards research.
[1] J. Allen, “Photoplethysmography and its application in clinical
physiology measurement,” Physiological Measurement, vol. A28, IoP
Publishing, 2007.J. Clerk Maxwell, A Treatise on Electricity and
Magnetism, 3rd ed., vol. 2. Oxford: Clarendon, 1892, pp.68–73.
[2] Zahoor-uddin, “Baseline Wandering Removal from Human
Electrocardiogram Signal using Projection Pursuit Gradient Ascent
Algorithm,” International Journal of Electrical & Computer Sciences
IJECS/IJENS, vol. 9, No:9, pp. 11-13.
[3] NHS Choices, http://www.nhs.uk/conditions/cardiovascular-
disease/pages/introduction.aspx [accessed 05/06/13].
[4] Jayasree, V K; Dr. Radhakrishnan, P. Selected cardiovascular studies
based on photoplethysmography technique. PhD Thesis.Cochin
University of Science and technology:India, Jan. 2009.
[5] Robert R. Stewart, “Review and A New F/K Analogue Design,”
Median Filtering, vol. 21, Dec. 1985, pp. 54-63.
[6] Physionet database, http://physionet.org/cgi-bin/atm/ATM [accessed
02/06/13].
[7] S. Nahar and M. ShahNoor bin Munir, "Automatic detection of
premature ventricular contraction beat using morphological
transformation and cross-correlation," in Signal Processing and
Communication Systems, 2009. ICSPCS 2009. 3rdInternational
Conference on, 2009, pp. 1-4.
[8] Matlab, http://www.mathworks.co.uk/help/matlab/ref/tstool.html
[accessed 07/08/13].
[9] Wikipedia, http://en.wikipedia.org/wiki/Root-mean-square_deviation
[accessed 07/08/13].

264
251